What the five-minute AI contract review changes for in-house legal skills

An Onit benchmark tested ten anonymised procurement contracts against senior-lawyer assessments. It reports an average review time of nearly 56 minutes for a junior lawyer and under five minutes for each of the tested large language models. The result is a vendor benchmark, not a general performance guarantee, but it clarifies what changes when routine review becomes fast.

When routine generation is no longer the bottleneck, the unit of legal work changes. Time moves from copying and summarising standard language toward framing the question, supplying the right context and testing the output against the commercial allocation of risk. The scarce skill is architectural judgment: knowing which assumptions matter, which exceptions require escalation and which parts of a document cannot be safely inferred.

The limit is accountability. AI generates text but does not exercise professional judgment, negotiate a commercial trade-off or stand accountable for what is filed. Incomplete context can produce a confident but wrong result. Human review is therefore not a ceremonial sign-off; it is the control that tests assumptions, catches omissions and owns the decision to accept or escalate the output.

For a legal team assessing an AI-enabled workflow, the relevant questions concern the operating model around the tool: how context is captured, how exceptions leave the standard path, what evidence a reviewer sees and who signs off on consequential risk allocation. Speed answers only the first question. The defensible measure is whether the workflow makes review boundaries and responsibility clear.